Bayesian inference in complex risk assessment with application to health hazards from abiotic toxin exposure

风险评估是一个迭代过程，旨在通过五个一般步骤来量化与一系列暴露条件对人类健康有关的危险：问题制定、危险识别、剂量-反应评估、暴露评估和风险表征。动物毒理学和人类观察性研究被用来表征毒素风险，但这些方法无法充分产生能够自信地描述真实世界条件和结果的函数。这些问题包括1)从动物到人类的推断；2)高剂量和低剂量反应之间的关系；3)平均、弹性和易感个体之间的反应差异；以及4)毒素之间相互作用对剂量-反应关系的影响。统计模型可以通过形式化关于复杂和非线性关系的假设以及模型中组件之间的相互作用来支持研究人员解释数据。数学模型与定量、真实世界和实验数据的结合对于研究生物学和生态学假说是至关重要的；然而，普遍缺乏评估非生物毒素暴露对人类的影响的系统方法，限制了研究人员估计与真实世界数据所知的环境毒素有关的个人健康风险的能力。我们的目标是研究贝叶斯方法，以开发一种方法来检查非生物毒素与健康结果之间的剂量-反应关系，考虑复杂的暴露数据、相互作用和结果。这项研究力求在贝叶斯推断和MCMC模拟研究进展的基础上，进一步表征剂量-反应关系及其相关的不确定性，为每个研究个体提供大量投入来模拟如此复杂的关系。目前推导剂量-反应关系的方法的局限性削弱了从环境危害中建立人类健康风险模型的适用性。投入的复杂性和影响敏感性的因素的高维性产生了一个超出了解决机械信息函数的范围的问题。即使是最常见的污染物之间的相互作用也会包含一个用传统测试方法需要数十年才能解决的组合问题，考虑到一个人每天接触的数千种非生物和生物暴露，以及个人的易感性特征，这一挑战是如此巨大，以至于目前的方法不太可能充分表征真正的风险。此外，随着生物系统复杂性的增加，模型的参数化问题变得更加重要和具有挑战性。考虑到影响个人易感性的投入范围，以及个人之间暴露和反应之间的复杂相互作用，建模框架必须能够详细了解促成健康结果的生态和生物过程。在将这种复杂模型参数化方面的发展需要进一步研究，以产生能够解释不同数据和结果来源的方法和技术。

Risk assessment is an iterative process that seeks to quantify the hazards associated with a range of exposure conditions on human health across five general steps: problem formulation, hazard identification, dose-response assessment, exposure assessment, and risk characterization. Animal toxicology and human observational studies are used to characterize toxin risk, but these methods fail to adequately produce a function that can confidently describe real-world conditions and outcomes. Among the concerns are 1) extrapolation from animals to humans; 2) the shape of the relationship between high-dose and low-dose response; 3) the difference in response between average, resilient, and susceptible individuals; and 4) the impact of interactions among toxins on the dose-response relationship. Statistical models can support researchers in interpreting data by formalizing hypotheses about complex and nonlinear relationships, as well as the interactions between components within a model. The integration of mathematical modelling, and quantitative, real-world and experimental data is essential to investigating biological and ecological hypotheses; however, a general scarcity of a systematic method for evaluating the impact, among humans, from abiotic toxin exposures remains, limiting researchers' capacity to estimate individual health risk associated with environmental toxins that are informed by real-world data. We aim to study Bayesian methods for the development of a methodology for examining the dose-response relationship between abiotic toxins and health outcomes, considering complex exposure data, interactions, and outcomes. This seeks to build on the developments in the study of Bayesian inference and MCMC simulations to further characterize the dose-response relationship and the associated uncertainty in modelling such a complex relationship with large amounts of inputs for each individual of study. The limitations in the current method of deriving a dose-response relationship undermine the applicability of human health risk modelling from environmental hazards. The complexity of inputs and the high dimensionality of factors influencing susceptibility produce a problem beyond the scope of solving a mechanistically-informed function. The interactions among even the most common pollutants would contain a combinatorial problem that would take decades to solve with conventional methods of testing, and given the thousands of abiotic and biotic exposures a person interacts with every day, alongside an individual's susceptibility profile, the enormity of this challenge is so great that it is unlikely that current methods will be able to adequately characterize true risk. Furthermore, the issue of parameterizing a model becomes more important and more challenging as the complexity of biological systems increase. Given the range of inputs that influence individual susceptibility and the complex interplay between exposure and response between individuals, modelling frameworks must be able to account for a detailed understanding of the ecological and biological processes that contribute to health outcomes. Development in the space of parameterizing such complex models requires further study to produce methodologies and techniques that can account for diverse sources of data and outcomes.